Uncertainty Quantification of a Deep Learning Based Fuel Property Prediction Model

Deep learning surrogate models for predicting properties of chemical compounds and mixtures have recently been shown to be promising for enabling data-driven novel fuel design and optimization, with the aim of improving efficiency and lowering emissions from combustion engines. However, given the low interpretability of typical neural network models, uncertainty quantification (UQ) is critical to ensure the reliability of predictions as well as the training datasets, and for a principled quantification of noise and its various sources. In this study, UQ of a multi-task deep learning model that simultaneously predicts the research octane number (RON), Motor Octane Number (MON), and Yield Sooting Index (YSI) of pure components and multicomponent blends, is performed. Both epistemic and aleatoric uncertainties are incorporated by utilizing various implementations of Monte Carlo Dropout, Bayesian Neural Network (BNN), and Gaussian Negative Log Likelihood (GNLL) loss function. A comparative analysis exploring these approaches is carried out to achieve the best trade-off between accuracy and calibration of the surrogate model.